GenAI adoption programme for a Top-20 pharma — from stalled pilots to governed production, with an Annex 22-aligned CoE

Pain points before the engagement

• GenAI pilots were multiplying across functions with no consistent governance or risk classification
• Use-cases that worked in demos stalled at production because Annex 22 posture was undefined
• Prompt and evaluation work was being re-done team by team — no shared library
• The HITL policy was a paragraph in a slide deck, not enforced in any tool
• Leadership could not see the AI portfolio in one place, let alone prioritise it

Problem

The pharma had gone from GenAI curiosity to GenAI proliferation in eighteen months. Every function had at least one pilot. None had a governance answer that survived contact with legal, security and quality. Leadership had funded a Centre of Excellence in principle but the operating model was undefined. The brief was to stand up a working AI CoE with an Annex 22-aligned assurance stack, and to move at least three high-value use-cases from pilot to production inside six months.

Approach

We stood up the CoE with a defined charter, staffing model and intake process. Use-cases were classified for risk on entry; low-risk use-cases could move fast, high-risk use-cases hit the assurance stack. A shared prompt and evaluation library ended the copy-paste-and-hope pattern across functions. HITL patterns were defined per risk class and enforced at the tool layer. Three flagship use-cases (regulatory intelligence, medical writing acceleration and CMC document review) were selected as the CoE's first production deliveries.